Post by Georg Grab

Excited to share that I've joined Prior Labs as Machine Learning Engineer! 🚀 While foundation models have revolutionized machine learning, most data science teams I’ve worked with in industry still predominantly use decade-old methods like XGBoost or GBDTs. Why? • Deep learning methods—especially transformer-based approaches—typically require enormous amounts of training data. • Many real-world business problems (fraud detection, risk assessment, recommender systems) have relatively limited labeled data. • Tree-based methods like XGBoost remain dominant because they work reasonably well even with small datasets. But these methods have real limitations: • They require substantial feature engineering to perform well. The real world is messy: missing values and distribution shifts need to be handled for every problem. • You train from scratch every time. Add a new feature to your dataset? Retrain the model. Prior Labs is bringing transformer-based methods to the ubiquitous world of tabular data with TabPFN. For the first time, traditional methods are being outperformed in benchmarks. No (re-)training, no feature engineering — just predictions. Incredibly excited to become part of the journey. The best place to learn about this method is the 2025 Nature publication: https://lnkd.in/du2JseRs Also, it’s open source! Check it out at: https://lnkd.in/dj6Gc_F9 Lastly, we’re hiring! Check out open positions here (or reach out to me): https://lnkd.in/d3G_B7ph